yoavg

Reinforcement Learning for Language Models

Yoav Goldberg, April 2023.

Why RL?

With the release of the ChatGPT model and followup large language models (LLMs), there was a lot of discussion of the importance of "RLHF training", that is, "reinforcement learning from human feedback". I was puzzled for a while as to why RL (Reinforcement Learning) is better than learning from demonstrations (a.k.a supervised learning) for training language models. Shouldn't learning from demonstrations (or, in language model terminology "instruction fine tuning", learning to immitate human written answers) be sufficient? I came up with a theoretical argument that was somewhat convincing. But I came to realize there is an additional argumment which not only supports the case of RL training, but also requires it, in particular for models like ChatGPT. This additional argument is spelled out in (the first half of) a talk by John Schulman from OpenAI. This post pretty much

tallguyjenks

# To permanently cache the credentials
git config --global credential.helper store

# To ignore files that could cause issues across different workspaces
touch .gitignore

echo ".obsidian/cache
.trash/
.DS_Store" > .gitignore

tangjeff0

'''
author:
    @tangjeff0
    https://www.notion.so/tangjeff0/Public-Home-0e2636bd409b454ea64079ad8213491f

inspired by: https://praxis.fortelabs.co/p-a-r-a-iii-building-an-idea-generator-400347ef3bb6/
with help from: https://medium.com/@jamiealexandre/introducing-notion-py-an-unofficial-python-api-wrapper-for-notion-so-603700f92369

credits:
    @jamiealexandre

dimitardanailov

Creating a Persistent Disk

Create a new instance

gcloud compute instances create gcelab --zone us-central1-c

Create a new persistent disk

weaming

#!/usr/bin/env python3
# coding: utf-8
"""
Author       : weaming
Created Time : 2018-05-26 21:32:59
Prerequisite:
    python3 -m pip install cson arrow
"""
import json
import os

ashokpant

#!/bin/bash

# install CUDA Toolkit v9.0
# instructions from https://developer.nvidia.com/cuda-downloads (linux -> x86_64 -> Ubuntu -> 16.04 -> deb)
CUDA_REPO_PKG="cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64-deb"
wget https://developer.nvidia.com/compute/cuda/9.0/Prod/local_installers/${CUDA_REPO_PKG}
sudo dpkg -i ${CUDA_REPO_PKG}
sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub
sudo apt-get update
sudo apt-get -y install cuda-9-0

rainsunny

Suppose you need to apply the same function to multiple columns in one DataFrame, one straight way is like this:

val newDF = oldDF.withColumn("colA", func("colA")).withColumn("colB", func("colB")).withColumn("colC", func("colC"))

If you want to save some type, you can try this:

1. Use select with varargs including *:

import spark.implicits._

MatthieuBizien

# coding: utf-8

import logging
import re
from collections import Counter

import numpy as np
import torch
from sklearn.datasets import fetch_20newsgroups
from torch.autograd import Variable

MichalZalecki

RUN ["apt-get", "update"]
RUN ["apt-get", "install", "-y", "zsh"]
RUN wget https://github.com/robbyrussell/oh-my-zsh/raw/master/tools/install.sh -O - | zsh || true

# docker exec -it my-app-container /bin/zsh

Swarchal

import pandas as pd
import numpy as np

def find_correlation(data, threshold=0.9, remove_negative=False):
    """
    Given a numeric pd.DataFrame, this will find highly correlated features,
    and return a list of features to remove.
    Parameters
    -----------
    data : pandas DataFrame